Comparison of Three Autoencoder-Based Models Using Molecular Dynamics (MD) Simulations Data

Molecular dynamics Molecular dynamics (MD) simulations have proven useful in studying the dynamics of Biomolecules, for instance proteins. However, the computational cost for conducting such simulations is high as reflected in the number of core-hours consumed in high performance computing (HPC) clusters. Although some techniques are available for enhancing the sampling of the conformational space, they usually make assumptions about the system by introducing empirical parameters. Machine learning (ML) models can overcome this issue because here, important features of the landscape can be inferred from the data themselves. In this work, we use an autoencoder ML model with three different flavors: Variational, Wasserstein, and Denoising to generate new protein conformations using MD trajectories as the training data. These generated structures can potentially enhance the ensemble of the original MD data.